Well the sample size should probably be mentioned as 22k / # students per class (~700 classes?), since the students results are purported to be correlated. The

Well the sample size should probably be mentioned as 22k / # students per class (~700 classes?), since the students results are purported to be correlated. The statistics is meagre as well. Regressing percentages, counts and dollars all non-normalised is bad form at the least. That said, I like the study. True random design takes guts and effort and gives data to be useful for years. We're still discussing RAND... On an important subject as well. Lot more to be found [1]. Thing that wonders me how state vs national differences are corrected for. On the one hand they shouldn't matter, since it is a random design. But the differences with national averages were huge. And far larger than the measured impact of the intervention. [1] https://www.povertyactionlab.org/evaluation/effects-kindergarten-classroom-earnings-united-states https://www.povertyactionlab.org/evaluation/effects-kinderga...